?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
KEYS: Fix up comments in key management code
[linux-2.6/next.git] / net / ipv4 / tcp.c
blob6c11eece262cf77a377a9fa901f4031c46745205
1 /*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Authors: Ross Biro
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 *
20 * Fixes:
21 * Alan Cox : Numerous verify_area() calls
22 * Alan Cox : Set the ACK bit on a reset
23 * Alan Cox : Stopped it crashing if it closed while
24 * sk->inuse=1 and was trying to connect
25 * (tcp_err()).
26 * Alan Cox : All icmp error handling was broken
27 * pointers passed where wrong and the
28 * socket was looked up backwards. Nobody
29 * tested any icmp error code obviously.
30 * Alan Cox : tcp_err() now handled properly. It
31 * wakes people on errors. poll
32 * behaves and the icmp error race
33 * has gone by moving it into sock.c
34 * Alan Cox : tcp_send_reset() fixed to work for
35 * everything not just packets for
36 * unknown sockets.
37 * Alan Cox : tcp option processing.
38 * Alan Cox : Reset tweaked (still not 100%) [Had
39 * syn rule wrong]
40 * Herp Rosmanith : More reset fixes
41 * Alan Cox : No longer acks invalid rst frames.
42 * Acking any kind of RST is right out.
43 * Alan Cox : Sets an ignore me flag on an rst
44 * receive otherwise odd bits of prattle
45 * escape still
46 * Alan Cox : Fixed another acking RST frame bug.
47 * Should stop LAN workplace lockups.
48 * Alan Cox : Some tidyups using the new skb list
49 * facilities
50 * Alan Cox : sk->keepopen now seems to work
51 * Alan Cox : Pulls options out correctly on accepts
52 * Alan Cox : Fixed assorted sk->rqueue->next errors
53 * Alan Cox : PSH doesn't end a TCP read. Switched a
54 * bit to skb ops.
55 * Alan Cox : Tidied tcp_data to avoid a potential
56 * nasty.
57 * Alan Cox : Added some better commenting, as the
58 * tcp is hard to follow
59 * Alan Cox : Removed incorrect check for 20 * psh
60 * Michael O'Reilly : ack < copied bug fix.
61 * Johannes Stille : Misc tcp fixes (not all in yet).
62 * Alan Cox : FIN with no memory -> CRASH
63 * Alan Cox : Added socket option proto entries.
64 * Also added awareness of them to accept.
65 * Alan Cox : Added TCP options (SOL_TCP)
66 * Alan Cox : Switched wakeup calls to callbacks,
67 * so the kernel can layer network
68 * sockets.
69 * Alan Cox : Use ip_tos/ip_ttl settings.
70 * Alan Cox : Handle FIN (more) properly (we hope).
71 * Alan Cox : RST frames sent on unsynchronised
72 * state ack error.
73 * Alan Cox : Put in missing check for SYN bit.
74 * Alan Cox : Added tcp_select_window() aka NET2E
75 * window non shrink trick.
76 * Alan Cox : Added a couple of small NET2E timer
77 * fixes
78 * Charles Hedrick : TCP fixes
79 * Toomas Tamm : TCP window fixes
80 * Alan Cox : Small URG fix to rlogin ^C ack fight
81 * Charles Hedrick : Rewrote most of it to actually work
82 * Linus : Rewrote tcp_read() and URG handling
83 * completely
84 * Gerhard Koerting: Fixed some missing timer handling
85 * Matthew Dillon : Reworked TCP machine states as per RFC
86 * Gerhard Koerting: PC/TCP workarounds
87 * Adam Caldwell : Assorted timer/timing errors
88 * Matthew Dillon : Fixed another RST bug
89 * Alan Cox : Move to kernel side addressing changes.
90 * Alan Cox : Beginning work on TCP fastpathing
91 * (not yet usable)
92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
93 * Alan Cox : TCP fast path debugging
94 * Alan Cox : Window clamping
95 * Michael Riepe : Bug in tcp_check()
96 * Matt Dillon : More TCP improvements and RST bug fixes
97 * Matt Dillon : Yet more small nasties remove from the
98 * TCP code (Be very nice to this man if
99 * tcp finally works 100%) 8)
100 * Alan Cox : BSD accept semantics.
101 * Alan Cox : Reset on closedown bug.
102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
103 * Michael Pall : Handle poll() after URG properly in
104 * all cases.
105 * Michael Pall : Undo the last fix in tcp_read_urg()
106 * (multi URG PUSH broke rlogin).
107 * Michael Pall : Fix the multi URG PUSH problem in
108 * tcp_readable(), poll() after URG
109 * works now.
110 * Michael Pall : recv(...,MSG_OOB) never blocks in the
111 * BSD api.
112 * Alan Cox : Changed the semantics of sk->socket to
113 * fix a race and a signal problem with
114 * accept() and async I/O.
115 * Alan Cox : Relaxed the rules on tcp_sendto().
116 * Yury Shevchuk : Really fixed accept() blocking problem.
117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
118 * clients/servers which listen in on
119 * fixed ports.
120 * Alan Cox : Cleaned the above up and shrank it to
121 * a sensible code size.
122 * Alan Cox : Self connect lockup fix.
123 * Alan Cox : No connect to multicast.
124 * Ross Biro : Close unaccepted children on master
125 * socket close.
126 * Alan Cox : Reset tracing code.
127 * Alan Cox : Spurious resets on shutdown.
128 * Alan Cox : Giant 15 minute/60 second timer error
129 * Alan Cox : Small whoops in polling before an
130 * accept.
131 * Alan Cox : Kept the state trace facility since
132 * it's handy for debugging.
133 * Alan Cox : More reset handler fixes.
134 * Alan Cox : Started rewriting the code based on
135 * the RFC's for other useful protocol
136 * references see: Comer, KA9Q NOS, and
137 * for a reference on the difference
138 * between specifications and how BSD
139 * works see the 4.4lite source.
140 * A.N.Kuznetsov : Don't time wait on completion of tidy
141 * close.
142 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
143 * Linus Torvalds : Fixed BSD port reuse to work first syn
144 * Alan Cox : Reimplemented timers as per the RFC
145 * and using multiple timers for sanity.
146 * Alan Cox : Small bug fixes, and a lot of new
147 * comments.
148 * Alan Cox : Fixed dual reader crash by locking
149 * the buffers (much like datagram.c)
150 * Alan Cox : Fixed stuck sockets in probe. A probe
151 * now gets fed up of retrying without
152 * (even a no space) answer.
153 * Alan Cox : Extracted closing code better
154 * Alan Cox : Fixed the closing state machine to
155 * resemble the RFC.
156 * Alan Cox : More 'per spec' fixes.
157 * Jorge Cwik : Even faster checksumming.
158 * Alan Cox : tcp_data() doesn't ack illegal PSH
159 * only frames. At least one pc tcp stack
160 * generates them.
161 * Alan Cox : Cache last socket.
162 * Alan Cox : Per route irtt.
163 * Matt Day : poll()->select() match BSD precisely on error
164 * Alan Cox : New buffers
165 * Marc Tamsky : Various sk->prot->retransmits and
166 * sk->retransmits misupdating fixed.
167 * Fixed tcp_write_timeout: stuck close,
168 * and TCP syn retries gets used now.
169 * Mark Yarvis : In tcp_read_wakeup(), don't send an
170 * ack if state is TCP_CLOSED.
171 * Alan Cox : Look up device on a retransmit - routes may
172 * change. Doesn't yet cope with MSS shrink right
173 * but it's a start!
174 * Marc Tamsky : Closing in closing fixes.
175 * Mike Shaver : RFC1122 verifications.
176 * Alan Cox : rcv_saddr errors.
177 * Alan Cox : Block double connect().
178 * Alan Cox : Small hooks for enSKIP.
179 * Alexey Kuznetsov: Path MTU discovery.
180 * Alan Cox : Support soft errors.
181 * Alan Cox : Fix MTU discovery pathological case
182 * when the remote claims no mtu!
183 * Marc Tamsky : TCP_CLOSE fix.
184 * Colin (G3TNE) : Send a reset on syn ack replies in
185 * window but wrong (fixes NT lpd problems)
186 * Pedro Roque : Better TCP window handling, delayed ack.
187 * Joerg Reuter : No modification of locked buffers in
188 * tcp_do_retransmit()
189 * Eric Schenk : Changed receiver side silly window
190 * avoidance algorithm to BSD style
191 * algorithm. This doubles throughput
192 * against machines running Solaris,
193 * and seems to result in general
194 * improvement.
195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
196 * Willy Konynenberg : Transparent proxying support.
197 * Mike McLagan : Routing by source
198 * Keith Owens : Do proper merging with partial SKB's in
199 * tcp_do_sendmsg to avoid burstiness.
200 * Eric Schenk : Fix fast close down bug with
201 * shutdown() followed by close().
202 * Andi Kleen : Make poll agree with SIGIO
203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
204 * lingertime == 0 (RFC 793 ABORT Call)
205 * Hirokazu Takahashi : Use copy_from_user() instead of
206 * csum_and_copy_from_user() if possible.
207 *
208 * This program is free software; you can redistribute it and/or
209 * modify it under the terms of the GNU General Public License
210 * as published by the Free Software Foundation; either version
211 * 2 of the License, or(at your option) any later version.
212 *
213 * Description of States:
214 *
215 * TCP_SYN_SENT sent a connection request, waiting for ack
216 *
217 * TCP_SYN_RECV received a connection request, sent ack,
218 * waiting for final ack in three-way handshake.
219 *
220 * TCP_ESTABLISHED connection established
221 *
222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
223 * transmission of remaining buffered data
224 *
225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
226 * to shutdown
227 *
228 * TCP_CLOSING both sides have shutdown but we still have
229 * data we have to finish sending
230 *
231 * TCP_TIME_WAIT timeout to catch resent junk before entering
232 * closed, can only be entered from FIN_WAIT2
233 * or CLOSING. Required because the other end
234 * may not have gotten our last ACK causing it
235 * to retransmit the data packet (which we ignore)
236 *
237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
238 * us to finish writing our data and to shutdown
239 * (we have to close() to move on to LAST_ACK)
240 *
241 * TCP_LAST_ACK out side has shutdown after remote has
242 * shutdown. There may still be data in our
243 * buffer that we have to finish sending
244 *
245 * TCP_CLOSE socket is finished
246 */
247
248 #include <linux/kernel.h>
249 #include <linux/module.h>
250 #include <linux/types.h>
251 #include <linux/fcntl.h>
252 #include <linux/poll.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/bootmem.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/crypto.h>
267 #include <linux/time.h>
268 #include <linux/slab.h>
269
270 #include <net/icmp.h>
271 #include <net/tcp.h>
272 #include <net/xfrm.h>
273 #include <net/ip.h>
274 #include <net/netdma.h>
275 #include <net/sock.h>
276
277 #include <asm/uaccess.h>
278 #include <asm/ioctls.h>
279
280 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
281
282 struct percpu_counter tcp_orphan_count;
283 EXPORT_SYMBOL_GPL(tcp_orphan_count);
284
285 long sysctl_tcp_mem[3] __read_mostly;
286 int sysctl_tcp_wmem[3] __read_mostly;
287 int sysctl_tcp_rmem[3] __read_mostly;
288
289 EXPORT_SYMBOL(sysctl_tcp_mem);
290 EXPORT_SYMBOL(sysctl_tcp_rmem);
291 EXPORT_SYMBOL(sysctl_tcp_wmem);
292
293 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
294 EXPORT_SYMBOL(tcp_memory_allocated);
295
296 /*
297 * Current number of TCP sockets.
298 */
299 struct percpu_counter tcp_sockets_allocated;
300 EXPORT_SYMBOL(tcp_sockets_allocated);
301
302 /*
303 * TCP splice context
304 */
305 struct tcp_splice_state {
306 struct pipe_inode_info *pipe;
307 size_t len;
308 unsigned int flags;
309 };
310
311 /*
312 * Pressure flag: try to collapse.
313 * Technical note: it is used by multiple contexts non atomically.
314 * All the __sk_mem_schedule() is of this nature: accounting
315 * is strict, actions are advisory and have some latency.
316 */
317 int tcp_memory_pressure __read_mostly;
318 EXPORT_SYMBOL(tcp_memory_pressure);
319
320 void tcp_enter_memory_pressure(struct sock *sk)
321 {
322 if (!tcp_memory_pressure) {
323 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
324 tcp_memory_pressure = 1;
325 }
326 }
327 EXPORT_SYMBOL(tcp_enter_memory_pressure);
328
329 /* Convert seconds to retransmits based on initial and max timeout */
330 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
331 {
332 u8 res = 0;
333
334 if (seconds > 0) {
335 int period = timeout;
336
337 res = 1;
338 while (seconds > period && res < 255) {
339 res++;
340 timeout <<= 1;
341 if (timeout > rto_max)
342 timeout = rto_max;
343 period += timeout;
344 }
345 }
346 return res;
347 }
348
349 /* Convert retransmits to seconds based on initial and max timeout */
350 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
351 {
352 int period = 0;
353
354 if (retrans > 0) {
355 period = timeout;
356 while (--retrans) {
357 timeout <<= 1;
358 if (timeout > rto_max)
359 timeout = rto_max;
360 period += timeout;
361 }
362 }
363 return period;
364 }
365
366 /*
367 * Wait for a TCP event.
368 *
369 * Note that we don't need to lock the socket, as the upper poll layers
370 * take care of normal races (between the test and the event) and we don't
371 * go look at any of the socket buffers directly.
372 */
373 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
374 {
375 unsigned int mask;
376 struct sock *sk = sock->sk;
377 struct tcp_sock *tp = tcp_sk(sk);
378
379 sock_poll_wait(file, sk_sleep(sk), wait);
380 if (sk->sk_state == TCP_LISTEN)
381 return inet_csk_listen_poll(sk);
382
383 /* Socket is not locked. We are protected from async events
384 * by poll logic and correct handling of state changes
385 * made by other threads is impossible in any case.
386 */
387
388 mask = 0;
389
390 /*
391 * POLLHUP is certainly not done right. But poll() doesn't
392 * have a notion of HUP in just one direction, and for a
393 * socket the read side is more interesting.
394 *
395 * Some poll() documentation says that POLLHUP is incompatible
396 * with the POLLOUT/POLLWR flags, so somebody should check this
397 * all. But careful, it tends to be safer to return too many
398 * bits than too few, and you can easily break real applications
399 * if you don't tell them that something has hung up!
400 *
401 * Check-me.
402 *
403 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
404 * our fs/select.c). It means that after we received EOF,
405 * poll always returns immediately, making impossible poll() on write()
406 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
407 * if and only if shutdown has been made in both directions.
408 * Actually, it is interesting to look how Solaris and DUX
409 * solve this dilemma. I would prefer, if POLLHUP were maskable,
410 * then we could set it on SND_SHUTDOWN. BTW examples given
411 * in Stevens' books assume exactly this behaviour, it explains
412 * why POLLHUP is incompatible with POLLOUT. --ANK
413 *
414 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
415 * blocking on fresh not-connected or disconnected socket. --ANK
416 */
417 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
418 mask |= POLLHUP;
419 if (sk->sk_shutdown & RCV_SHUTDOWN)
420 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
421
422 /* Connected? */
423 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
424 int target = sock_rcvlowat(sk, 0, INT_MAX);
425
426 if (tp->urg_seq == tp->copied_seq &&
427 !sock_flag(sk, SOCK_URGINLINE) &&
428 tp->urg_data)
429 target++;
430
431 /* Potential race condition. If read of tp below will
432 * escape above sk->sk_state, we can be illegally awaken
433 * in SYN_* states. */
434 if (tp->rcv_nxt - tp->copied_seq >= target)
435 mask |= POLLIN | POLLRDNORM;
436
437 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
438 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
439 mask |= POLLOUT | POLLWRNORM;
440 } else { /* send SIGIO later */
441 set_bit(SOCK_ASYNC_NOSPACE,
442 &sk->sk_socket->flags);
443 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
444
445 /* Race breaker. If space is freed after
446 * wspace test but before the flags are set,
447 * IO signal will be lost.
448 */
449 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
450 mask |= POLLOUT | POLLWRNORM;
451 }
452 } else
453 mask |= POLLOUT | POLLWRNORM;
454
455 if (tp->urg_data & TCP_URG_VALID)
456 mask |= POLLPRI;
457 }
458 /* This barrier is coupled with smp_wmb() in tcp_reset() */
459 smp_rmb();
460 if (sk->sk_err)
461 mask |= POLLERR;
462
463 return mask;
464 }
465 EXPORT_SYMBOL(tcp_poll);
466
467 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
468 {
469 struct tcp_sock *tp = tcp_sk(sk);
470 int answ;
471
472 switch (cmd) {
473 case SIOCINQ:
474 if (sk->sk_state == TCP_LISTEN)
475 return -EINVAL;
476
477 lock_sock(sk);
478 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
479 answ = 0;
480 else if (sock_flag(sk, SOCK_URGINLINE) ||
481 !tp->urg_data ||
482 before(tp->urg_seq, tp->copied_seq) ||
483 !before(tp->urg_seq, tp->rcv_nxt)) {
484 struct sk_buff *skb;
485
486 answ = tp->rcv_nxt - tp->copied_seq;
487
488 /* Subtract 1, if FIN is in queue. */
489 skb = skb_peek_tail(&sk->sk_receive_queue);
490 if (answ && skb)
491 answ -= tcp_hdr(skb)->fin;
492 } else
493 answ = tp->urg_seq - tp->copied_seq;
494 release_sock(sk);
495 break;
496 case SIOCATMARK:
497 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
498 break;
499 case SIOCOUTQ:
500 if (sk->sk_state == TCP_LISTEN)
501 return -EINVAL;
502
503 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
504 answ = 0;
505 else
506 answ = tp->write_seq - tp->snd_una;
507 break;
508 default:
509 return -ENOIOCTLCMD;
510 }
511
512 return put_user(answ, (int __user *)arg);
513 }
514 EXPORT_SYMBOL(tcp_ioctl);
515
516 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
517 {
518 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
519 tp->pushed_seq = tp->write_seq;
520 }
521
522 static inline int forced_push(struct tcp_sock *tp)
523 {
524 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
525 }
526
527 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
528 {
529 struct tcp_sock *tp = tcp_sk(sk);
530 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
531
532 skb->csum = 0;
533 tcb->seq = tcb->end_seq = tp->write_seq;
534 tcb->flags = TCPHDR_ACK;
535 tcb->sacked = 0;
536 skb_header_release(skb);
537 tcp_add_write_queue_tail(sk, skb);
538 sk->sk_wmem_queued += skb->truesize;
539 sk_mem_charge(sk, skb->truesize);
540 if (tp->nonagle & TCP_NAGLE_PUSH)
541 tp->nonagle &= ~TCP_NAGLE_PUSH;
542 }
543
544 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
545 {
546 if (flags & MSG_OOB)
547 tp->snd_up = tp->write_seq;
548 }
549
550 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
551 int nonagle)
552 {
553 if (tcp_send_head(sk)) {
554 struct tcp_sock *tp = tcp_sk(sk);
555
556 if (!(flags & MSG_MORE) || forced_push(tp))
557 tcp_mark_push(tp, tcp_write_queue_tail(sk));
558
559 tcp_mark_urg(tp, flags);
560 __tcp_push_pending_frames(sk, mss_now,
561 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
562 }
563 }
564
565 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
566 unsigned int offset, size_t len)
567 {
568 struct tcp_splice_state *tss = rd_desc->arg.data;
569 int ret;
570
571 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
572 tss->flags);
573 if (ret > 0)
574 rd_desc->count -= ret;
575 return ret;
576 }
577
578 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
579 {
580 /* Store TCP splice context information in read_descriptor_t. */
581 read_descriptor_t rd_desc = {
582 .arg.data = tss,
583 .count = tss->len,
584 };
585
586 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
587 }
588
589 /**
590 * tcp_splice_read - splice data from TCP socket to a pipe
591 * @sock: socket to splice from
592 * @ppos: position (not valid)
593 * @pipe: pipe to splice to
594 * @len: number of bytes to splice
595 * @flags: splice modifier flags
596 *
597 * Description:
598 * Will read pages from given socket and fill them into a pipe.
599 *
600 **/
601 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
602 struct pipe_inode_info *pipe, size_t len,
603 unsigned int flags)
604 {
605 struct sock *sk = sock->sk;
606 struct tcp_splice_state tss = {
607 .pipe = pipe,
608 .len = len,
609 .flags = flags,
610 };
611 long timeo;
612 ssize_t spliced;
613 int ret;
614
615 sock_rps_record_flow(sk);
616 /*
617 * We can't seek on a socket input
618 */
619 if (unlikely(*ppos))
620 return -ESPIPE;
621
622 ret = spliced = 0;
623
624 lock_sock(sk);
625
626 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
627 while (tss.len) {
628 ret = __tcp_splice_read(sk, &tss);
629 if (ret < 0)
630 break;
631 else if (!ret) {
632 if (spliced)
633 break;
634 if (sock_flag(sk, SOCK_DONE))
635 break;
636 if (sk->sk_err) {
637 ret = sock_error(sk);
638 break;
639 }
640 if (sk->sk_shutdown & RCV_SHUTDOWN)
641 break;
642 if (sk->sk_state == TCP_CLOSE) {
643 /*
644 * This occurs when user tries to read
645 * from never connected socket.
646 */
647 if (!sock_flag(sk, SOCK_DONE))
648 ret = -ENOTCONN;
649 break;
650 }
651 if (!timeo) {
652 ret = -EAGAIN;
653 break;
654 }
655 sk_wait_data(sk, &timeo);
656 if (signal_pending(current)) {
657 ret = sock_intr_errno(timeo);
658 break;
659 }
660 continue;
661 }
662 tss.len -= ret;
663 spliced += ret;
664
665 if (!timeo)
666 break;
667 release_sock(sk);
668 lock_sock(sk);
669
670 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
671 (sk->sk_shutdown & RCV_SHUTDOWN) ||
672 signal_pending(current))
673 break;
674 }
675
676 release_sock(sk);
677
678 if (spliced)
679 return spliced;
680
681 return ret;
682 }
683 EXPORT_SYMBOL(tcp_splice_read);
684
685 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
686 {
687 struct sk_buff *skb;
688
689 /* The TCP header must be at least 32-bit aligned. */
690 size = ALIGN(size, 4);
691
692 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
693 if (skb) {
694 if (sk_wmem_schedule(sk, skb->truesize)) {
695 /*
696 * Make sure that we have exactly size bytes
697 * available to the caller, no more, no less.
698 */
699 skb_reserve(skb, skb_tailroom(skb) - size);
700 return skb;
701 }
702 __kfree_skb(skb);
703 } else {
704 sk->sk_prot->enter_memory_pressure(sk);
705 sk_stream_moderate_sndbuf(sk);
706 }
707 return NULL;
708 }
709
710 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
711 int large_allowed)
712 {
713 struct tcp_sock *tp = tcp_sk(sk);
714 u32 xmit_size_goal, old_size_goal;
715
716 xmit_size_goal = mss_now;
717
718 if (large_allowed && sk_can_gso(sk)) {
719 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
720 inet_csk(sk)->icsk_af_ops->net_header_len -
721 inet_csk(sk)->icsk_ext_hdr_len -
722 tp->tcp_header_len);
723
724 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
725
726 /* We try hard to avoid divides here */
727 old_size_goal = tp->xmit_size_goal_segs * mss_now;
728
729 if (likely(old_size_goal <= xmit_size_goal &&
730 old_size_goal + mss_now > xmit_size_goal)) {
731 xmit_size_goal = old_size_goal;
732 } else {
733 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
734 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
735 }
736 }
737
738 return max(xmit_size_goal, mss_now);
739 }
740
741 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
742 {
743 int mss_now;
744
745 mss_now = tcp_current_mss(sk);
746 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
747
748 return mss_now;
749 }
750
751 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
752 size_t psize, int flags)
753 {
754 struct tcp_sock *tp = tcp_sk(sk);
755 int mss_now, size_goal;
756 int err;
757 ssize_t copied;
758 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
759
760 /* Wait for a connection to finish. */
761 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
762 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
763 goto out_err;
764
765 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
766
767 mss_now = tcp_send_mss(sk, &size_goal, flags);
768 copied = 0;
769
770 err = -EPIPE;
771 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
772 goto out_err;
773
774 while (psize > 0) {
775 struct sk_buff *skb = tcp_write_queue_tail(sk);
776 struct page *page = pages[poffset / PAGE_SIZE];
777 int copy, i, can_coalesce;
778 int offset = poffset % PAGE_SIZE;
779 int size = min_t(size_t, psize, PAGE_SIZE - offset);
780
781 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
782 new_segment:
783 if (!sk_stream_memory_free(sk))
784 goto wait_for_sndbuf;
785
786 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
787 if (!skb)
788 goto wait_for_memory;
789
790 skb_entail(sk, skb);
791 copy = size_goal;
792 }
793
794 if (copy > size)
795 copy = size;
796
797 i = skb_shinfo(skb)->nr_frags;
798 can_coalesce = skb_can_coalesce(skb, i, page, offset);
799 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
800 tcp_mark_push(tp, skb);
801 goto new_segment;
802 }
803 if (!sk_wmem_schedule(sk, copy))
804 goto wait_for_memory;
805
806 if (can_coalesce) {
807 skb_shinfo(skb)->frags[i - 1].size += copy;
808 } else {
809 get_page(page);
810 skb_fill_page_desc(skb, i, page, offset, copy);
811 }
812
813 skb->len += copy;
814 skb->data_len += copy;
815 skb->truesize += copy;
816 sk->sk_wmem_queued += copy;
817 sk_mem_charge(sk, copy);
818 skb->ip_summed = CHECKSUM_PARTIAL;
819 tp->write_seq += copy;
820 TCP_SKB_CB(skb)->end_seq += copy;
821 skb_shinfo(skb)->gso_segs = 0;
822
823 if (!copied)
824 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
825
826 copied += copy;
827 poffset += copy;
828 if (!(psize -= copy))
829 goto out;
830
831 if (skb->len < size_goal || (flags & MSG_OOB))
832 continue;
833
834 if (forced_push(tp)) {
835 tcp_mark_push(tp, skb);
836 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
837 } else if (skb == tcp_send_head(sk))
838 tcp_push_one(sk, mss_now);
839 continue;
840
841 wait_for_sndbuf:
842 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
843 wait_for_memory:
844 if (copied)
845 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
846
847 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
848 goto do_error;
849
850 mss_now = tcp_send_mss(sk, &size_goal, flags);
851 }
852
853 out:
854 if (copied)
855 tcp_push(sk, flags, mss_now, tp->nonagle);
856 return copied;
857
858 do_error:
859 if (copied)
860 goto out;
861 out_err:
862 return sk_stream_error(sk, flags, err);
863 }
864
865 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
866 size_t size, int flags)
867 {
868 ssize_t res;
869
870 if (!(sk->sk_route_caps & NETIF_F_SG) ||
871 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
872 return sock_no_sendpage(sk->sk_socket, page, offset, size,
873 flags);
874
875 lock_sock(sk);
876 TCP_CHECK_TIMER(sk);
877 res = do_tcp_sendpages(sk, &page, offset, size, flags);
878 TCP_CHECK_TIMER(sk);
879 release_sock(sk);
880 return res;
881 }
882 EXPORT_SYMBOL(tcp_sendpage);
883
884 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
885 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
886
887 static inline int select_size(struct sock *sk, int sg)
888 {
889 struct tcp_sock *tp = tcp_sk(sk);
890 int tmp = tp->mss_cache;
891
892 if (sg) {
893 if (sk_can_gso(sk))
894 tmp = 0;
895 else {
896 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
897
898 if (tmp >= pgbreak &&
899 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
900 tmp = pgbreak;
901 }
902 }
903
904 return tmp;
905 }
906
907 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
908 size_t size)
909 {
910 struct iovec *iov;
911 struct tcp_sock *tp = tcp_sk(sk);
912 struct sk_buff *skb;
913 int iovlen, flags;
914 int mss_now, size_goal;
915 int sg, err, copied;
916 long timeo;
917
918 lock_sock(sk);
919 TCP_CHECK_TIMER(sk);
920
921 flags = msg->msg_flags;
922 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
923
924 /* Wait for a connection to finish. */
925 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
926 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
927 goto out_err;
928
929 /* This should be in poll */
930 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
931
932 mss_now = tcp_send_mss(sk, &size_goal, flags);
933
934 /* Ok commence sending. */
935 iovlen = msg->msg_iovlen;
936 iov = msg->msg_iov;
937 copied = 0;
938
939 err = -EPIPE;
940 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
941 goto out_err;
942
943 sg = sk->sk_route_caps & NETIF_F_SG;
944
945 while (--iovlen >= 0) {
946 size_t seglen = iov->iov_len;
947 unsigned char __user *from = iov->iov_base;
948
949 iov++;
950
951 while (seglen > 0) {
952 int copy = 0;
953 int max = size_goal;
954
955 skb = tcp_write_queue_tail(sk);
956 if (tcp_send_head(sk)) {
957 if (skb->ip_summed == CHECKSUM_NONE)
958 max = mss_now;
959 copy = max - skb->len;
960 }
961
962 if (copy <= 0) {
963 new_segment:
964 /* Allocate new segment. If the interface is SG,
965 * allocate skb fitting to single page.
966 */
967 if (!sk_stream_memory_free(sk))
968 goto wait_for_sndbuf;
969
970 skb = sk_stream_alloc_skb(sk,
971 select_size(sk, sg),
972 sk->sk_allocation);
973 if (!skb)
974 goto wait_for_memory;
975
976 /*
977 * Check whether we can use HW checksum.
978 */
979 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
980 skb->ip_summed = CHECKSUM_PARTIAL;
981
982 skb_entail(sk, skb);
983 copy = size_goal;
984 max = size_goal;
985 }
986
987 /* Try to append data to the end of skb. */
988 if (copy > seglen)
989 copy = seglen;
990
991 /* Where to copy to? */
992 if (skb_tailroom(skb) > 0) {
993 /* We have some space in skb head. Superb! */
994 if (copy > skb_tailroom(skb))
995 copy = skb_tailroom(skb);
996 if ((err = skb_add_data(skb, from, copy)) != 0)
997 goto do_fault;
998 } else {
999 int merge = 0;
1000 int i = skb_shinfo(skb)->nr_frags;
1001 struct page *page = TCP_PAGE(sk);
1002 int off = TCP_OFF(sk);
1003
1004 if (skb_can_coalesce(skb, i, page, off) &&
1005 off != PAGE_SIZE) {
1006 /* We can extend the last page
1007 * fragment. */
1008 merge = 1;
1009 } else if (i == MAX_SKB_FRAGS || !sg) {
1010 /* Need to add new fragment and cannot
1011 * do this because interface is non-SG,
1012 * or because all the page slots are
1013 * busy. */
1014 tcp_mark_push(tp, skb);
1015 goto new_segment;
1016 } else if (page) {
1017 if (off == PAGE_SIZE) {
1018 put_page(page);
1019 TCP_PAGE(sk) = page = NULL;
1020 off = 0;
1021 }
1022 } else
1023 off = 0;
1024
1025 if (copy > PAGE_SIZE - off)
1026 copy = PAGE_SIZE - off;
1027
1028 if (!sk_wmem_schedule(sk, copy))
1029 goto wait_for_memory;
1030
1031 if (!page) {
1032 /* Allocate new cache page. */
1033 if (!(page = sk_stream_alloc_page(sk)))
1034 goto wait_for_memory;
1035 }
1036
1037 /* Time to copy data. We are close to
1038 * the end! */
1039 err = skb_copy_to_page(sk, from, skb, page,
1040 off, copy);
1041 if (err) {
1042 /* If this page was new, give it to the
1043 * socket so it does not get leaked.
1044 */
1045 if (!TCP_PAGE(sk)) {
1046 TCP_PAGE(sk) = page;
1047 TCP_OFF(sk) = 0;
1048 }
1049 goto do_error;
1050 }
1051
1052 /* Update the skb. */
1053 if (merge) {
1054 skb_shinfo(skb)->frags[i - 1].size +=
1055 copy;
1056 } else {
1057 skb_fill_page_desc(skb, i, page, off, copy);
1058 if (TCP_PAGE(sk)) {
1059 get_page(page);
1060 } else if (off + copy < PAGE_SIZE) {
1061 get_page(page);
1062 TCP_PAGE(sk) = page;
1063 }
1064 }
1065
1066 TCP_OFF(sk) = off + copy;
1067 }
1068
1069 if (!copied)
1070 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
1071
1072 tp->write_seq += copy;
1073 TCP_SKB_CB(skb)->end_seq += copy;
1074 skb_shinfo(skb)->gso_segs = 0;
1075
1076 from += copy;
1077 copied += copy;
1078 if ((seglen -= copy) == 0 && iovlen == 0)
1079 goto out;
1080
1081 if (skb->len < max || (flags & MSG_OOB))
1082 continue;
1083
1084 if (forced_push(tp)) {
1085 tcp_mark_push(tp, skb);
1086 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1087 } else if (skb == tcp_send_head(sk))
1088 tcp_push_one(sk, mss_now);
1089 continue;
1090
1091 wait_for_sndbuf:
1092 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1093 wait_for_memory:
1094 if (copied)
1095 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1096
1097 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1098 goto do_error;
1099
1100 mss_now = tcp_send_mss(sk, &size_goal, flags);
1101 }
1102 }
1103
1104 out:
1105 if (copied)
1106 tcp_push(sk, flags, mss_now, tp->nonagle);
1107 TCP_CHECK_TIMER(sk);
1108 release_sock(sk);
1109 return copied;
1110
1111 do_fault:
1112 if (!skb->len) {
1113 tcp_unlink_write_queue(skb, sk);
1114 /* It is the one place in all of TCP, except connection
1115 * reset, where we can be unlinking the send_head.
1116 */
1117 tcp_check_send_head(sk, skb);
1118 sk_wmem_free_skb(sk, skb);
1119 }
1120
1121 do_error:
1122 if (copied)
1123 goto out;
1124 out_err:
1125 err = sk_stream_error(sk, flags, err);
1126 TCP_CHECK_TIMER(sk);
1127 release_sock(sk);
1128 return err;
1129 }
1130 EXPORT_SYMBOL(tcp_sendmsg);
1131
1132 /*
1133 * Handle reading urgent data. BSD has very simple semantics for
1134 * this, no blocking and very strange errors 8)
1135 */
1136
1137 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1138 {
1139 struct tcp_sock *tp = tcp_sk(sk);
1140
1141 /* No URG data to read. */
1142 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1143 tp->urg_data == TCP_URG_READ)
1144 return -EINVAL; /* Yes this is right ! */
1145
1146 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1147 return -ENOTCONN;
1148
1149 if (tp->urg_data & TCP_URG_VALID) {
1150 int err = 0;
1151 char c = tp->urg_data;
1152
1153 if (!(flags & MSG_PEEK))
1154 tp->urg_data = TCP_URG_READ;
1155
1156 /* Read urgent data. */
1157 msg->msg_flags |= MSG_OOB;
1158
1159 if (len > 0) {
1160 if (!(flags & MSG_TRUNC))
1161 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1162 len = 1;
1163 } else
1164 msg->msg_flags |= MSG_TRUNC;
1165
1166 return err ? -EFAULT : len;
1167 }
1168
1169 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1170 return 0;
1171
1172 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1173 * the available implementations agree in this case:
1174 * this call should never block, independent of the
1175 * blocking state of the socket.
1176 * Mike <pall@rz.uni-karlsruhe.de>
1177 */
1178 return -EAGAIN;
1179 }
1180
1181 /* Clean up the receive buffer for full frames taken by the user,
1182 * then send an ACK if necessary. COPIED is the number of bytes
1183 * tcp_recvmsg has given to the user so far, it speeds up the
1184 * calculation of whether or not we must ACK for the sake of
1185 * a window update.
1186 */
1187 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1188 {
1189 struct tcp_sock *tp = tcp_sk(sk);
1190 int time_to_ack = 0;
1191
1192 #if TCP_DEBUG
1193 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1194
1195 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1196 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1197 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1198 #endif
1199
1200 if (inet_csk_ack_scheduled(sk)) {
1201 const struct inet_connection_sock *icsk = inet_csk(sk);
1202 /* Delayed ACKs frequently hit locked sockets during bulk
1203 * receive. */
1204 if (icsk->icsk_ack.blocked ||
1205 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1206 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1207 /*
1208 * If this read emptied read buffer, we send ACK, if
1209 * connection is not bidirectional, user drained
1210 * receive buffer and there was a small segment
1211 * in queue.
1212 */
1213 (copied > 0 &&
1214 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1215 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1216 !icsk->icsk_ack.pingpong)) &&
1217 !atomic_read(&sk->sk_rmem_alloc)))
1218 time_to_ack = 1;
1219 }
1220
1221 /* We send an ACK if we can now advertise a non-zero window
1222 * which has been raised "significantly".
1223 *
1224 * Even if window raised up to infinity, do not send window open ACK
1225 * in states, where we will not receive more. It is useless.
1226 */
1227 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1228 __u32 rcv_window_now = tcp_receive_window(tp);
1229
1230 /* Optimize, __tcp_select_window() is not cheap. */
1231 if (2*rcv_window_now <= tp->window_clamp) {
1232 __u32 new_window = __tcp_select_window(sk);
1233
1234 /* Send ACK now, if this read freed lots of space
1235 * in our buffer. Certainly, new_window is new window.
1236 * We can advertise it now, if it is not less than current one.
1237 * "Lots" means "at least twice" here.
1238 */
1239 if (new_window && new_window >= 2 * rcv_window_now)
1240 time_to_ack = 1;
1241 }
1242 }
1243 if (time_to_ack)
1244 tcp_send_ack(sk);
1245 }
1246
1247 static void tcp_prequeue_process(struct sock *sk)
1248 {
1249 struct sk_buff *skb;
1250 struct tcp_sock *tp = tcp_sk(sk);
1251
1252 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1253
1254 /* RX process wants to run with disabled BHs, though it is not
1255 * necessary */
1256 local_bh_disable();
1257 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1258 sk_backlog_rcv(sk, skb);
1259 local_bh_enable();
1260
1261 /* Clear memory counter. */
1262 tp->ucopy.memory = 0;
1263 }
1264
1265 #ifdef CONFIG_NET_DMA
1266 static void tcp_service_net_dma(struct sock *sk, bool wait)
1267 {
1268 dma_cookie_t done, used;
1269 dma_cookie_t last_issued;
1270 struct tcp_sock *tp = tcp_sk(sk);
1271
1272 if (!tp->ucopy.dma_chan)
1273 return;
1274
1275 last_issued = tp->ucopy.dma_cookie;
1276 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1277
1278 do {
1279 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1280 last_issued, &done,
1281 &used) == DMA_SUCCESS) {
1282 /* Safe to free early-copied skbs now */
1283 __skb_queue_purge(&sk->sk_async_wait_queue);
1284 break;
1285 } else {
1286 struct sk_buff *skb;
1287 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1288 (dma_async_is_complete(skb->dma_cookie, done,
1289 used) == DMA_SUCCESS)) {
1290 __skb_dequeue(&sk->sk_async_wait_queue);
1291 kfree_skb(skb);
1292 }
1293 }
1294 } while (wait);
1295 }
1296 #endif
1297
1298 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1299 {
1300 struct sk_buff *skb;
1301 u32 offset;
1302
1303 skb_queue_walk(&sk->sk_receive_queue, skb) {
1304 offset = seq - TCP_SKB_CB(skb)->seq;
1305 if (tcp_hdr(skb)->syn)
1306 offset--;
1307 if (offset < skb->len || tcp_hdr(skb)->fin) {
1308 *off = offset;
1309 return skb;
1310 }
1311 }
1312 return NULL;
1313 }
1314
1315 /*
1316 * This routine provides an alternative to tcp_recvmsg() for routines
1317 * that would like to handle copying from skbuffs directly in 'sendfile'
1318 * fashion.
1319 * Note:
1320 * - It is assumed that the socket was locked by the caller.
1321 * - The routine does not block.
1322 * - At present, there is no support for reading OOB data
1323 * or for 'peeking' the socket using this routine
1324 * (although both would be easy to implement).
1325 */
1326 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1327 sk_read_actor_t recv_actor)
1328 {
1329 struct sk_buff *skb;
1330 struct tcp_sock *tp = tcp_sk(sk);
1331 u32 seq = tp->copied_seq;
1332 u32 offset;
1333 int copied = 0;
1334
1335 if (sk->sk_state == TCP_LISTEN)
1336 return -ENOTCONN;
1337 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1338 if (offset < skb->len) {
1339 int used;
1340 size_t len;
1341
1342 len = skb->len - offset;
1343 /* Stop reading if we hit a patch of urgent data */
1344 if (tp->urg_data) {
1345 u32 urg_offset = tp->urg_seq - seq;
1346 if (urg_offset < len)
1347 len = urg_offset;
1348 if (!len)
1349 break;
1350 }
1351 used = recv_actor(desc, skb, offset, len);
1352 if (used < 0) {
1353 if (!copied)
1354 copied = used;
1355 break;
1356 } else if (used <= len) {
1357 seq += used;
1358 copied += used;
1359 offset += used;
1360 }
1361 /*
1362 * If recv_actor drops the lock (e.g. TCP splice
1363 * receive) the skb pointer might be invalid when
1364 * getting here: tcp_collapse might have deleted it
1365 * while aggregating skbs from the socket queue.
1366 */
1367 skb = tcp_recv_skb(sk, seq-1, &offset);
1368 if (!skb || (offset+1 != skb->len))
1369 break;
1370 }
1371 if (tcp_hdr(skb)->fin) {
1372 sk_eat_skb(sk, skb, 0);
1373 ++seq;
1374 break;
1375 }
1376 sk_eat_skb(sk, skb, 0);
1377 if (!desc->count)
1378 break;
1379 tp->copied_seq = seq;
1380 }
1381 tp->copied_seq = seq;
1382
1383 tcp_rcv_space_adjust(sk);
1384
1385 /* Clean up data we have read: This will do ACK frames. */
1386 if (copied > 0)
1387 tcp_cleanup_rbuf(sk, copied);
1388 return copied;
1389 }
1390 EXPORT_SYMBOL(tcp_read_sock);
1391
1392 /*
1393 * This routine copies from a sock struct into the user buffer.
1394 *
1395 * Technical note: in 2.3 we work on _locked_ socket, so that
1396 * tricks with *seq access order and skb->users are not required.
1397 * Probably, code can be easily improved even more.
1398 */
1399
1400 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1401 size_t len, int nonblock, int flags, int *addr_len)
1402 {
1403 struct tcp_sock *tp = tcp_sk(sk);
1404 int copied = 0;
1405 u32 peek_seq;
1406 u32 *seq;
1407 unsigned long used;
1408 int err;
1409 int target; /* Read at least this many bytes */
1410 long timeo;
1411 struct task_struct *user_recv = NULL;
1412 int copied_early = 0;
1413 struct sk_buff *skb;
1414 u32 urg_hole = 0;
1415
1416 lock_sock(sk);
1417
1418 TCP_CHECK_TIMER(sk);
1419
1420 err = -ENOTCONN;
1421 if (sk->sk_state == TCP_LISTEN)
1422 goto out;
1423
1424 timeo = sock_rcvtimeo(sk, nonblock);
1425
1426 /* Urgent data needs to be handled specially. */
1427 if (flags & MSG_OOB)
1428 goto recv_urg;
1429
1430 seq = &tp->copied_seq;
1431 if (flags & MSG_PEEK) {
1432 peek_seq = tp->copied_seq;
1433 seq = &peek_seq;
1434 }
1435
1436 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1437
1438 #ifdef CONFIG_NET_DMA
1439 tp->ucopy.dma_chan = NULL;
1440 preempt_disable();
1441 skb = skb_peek_tail(&sk->sk_receive_queue);
1442 {
1443 int available = 0;
1444
1445 if (skb)
1446 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1447 if ((available < target) &&
1448 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1449 !sysctl_tcp_low_latency &&
1450 dma_find_channel(DMA_MEMCPY)) {
1451 preempt_enable_no_resched();
1452 tp->ucopy.pinned_list =
1453 dma_pin_iovec_pages(msg->msg_iov, len);
1454 } else {
1455 preempt_enable_no_resched();
1456 }
1457 }
1458 #endif
1459
1460 do {
1461 u32 offset;
1462
1463 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1464 if (tp->urg_data && tp->urg_seq == *seq) {
1465 if (copied)
1466 break;
1467 if (signal_pending(current)) {
1468 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1469 break;
1470 }
1471 }
1472
1473 /* Next get a buffer. */
1474
1475 skb_queue_walk(&sk->sk_receive_queue, skb) {
1476 /* Now that we have two receive queues this
1477 * shouldn't happen.
1478 */
1479 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1480 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1481 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1482 flags))
1483 break;
1484
1485 offset = *seq - TCP_SKB_CB(skb)->seq;
1486 if (tcp_hdr(skb)->syn)
1487 offset--;
1488 if (offset < skb->len)
1489 goto found_ok_skb;
1490 if (tcp_hdr(skb)->fin)
1491 goto found_fin_ok;
1492 WARN(!(flags & MSG_PEEK),
1493 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1494 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1495 }
1496
1497 /* Well, if we have backlog, try to process it now yet. */
1498
1499 if (copied >= target && !sk->sk_backlog.tail)
1500 break;
1501
1502 if (copied) {
1503 if (sk->sk_err ||
1504 sk->sk_state == TCP_CLOSE ||
1505 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1506 !timeo ||
1507 signal_pending(current))
1508 break;
1509 } else {
1510 if (sock_flag(sk, SOCK_DONE))
1511 break;
1512
1513 if (sk->sk_err) {
1514 copied = sock_error(sk);
1515 break;
1516 }
1517
1518 if (sk->sk_shutdown & RCV_SHUTDOWN)
1519 break;
1520
1521 if (sk->sk_state == TCP_CLOSE) {
1522 if (!sock_flag(sk, SOCK_DONE)) {
1523 /* This occurs when user tries to read
1524 * from never connected socket.
1525 */
1526 copied = -ENOTCONN;
1527 break;
1528 }
1529 break;
1530 }
1531
1532 if (!timeo) {
1533 copied = -EAGAIN;
1534 break;
1535 }
1536
1537 if (signal_pending(current)) {
1538 copied = sock_intr_errno(timeo);
1539 break;
1540 }
1541 }
1542
1543 tcp_cleanup_rbuf(sk, copied);
1544
1545 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1546 /* Install new reader */
1547 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1548 user_recv = current;
1549 tp->ucopy.task = user_recv;
1550 tp->ucopy.iov = msg->msg_iov;
1551 }
1552
1553 tp->ucopy.len = len;
1554
1555 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1556 !(flags & (MSG_PEEK | MSG_TRUNC)));
1557
1558 /* Ugly... If prequeue is not empty, we have to
1559 * process it before releasing socket, otherwise
1560 * order will be broken at second iteration.
1561 * More elegant solution is required!!!
1562 *
1563 * Look: we have the following (pseudo)queues:
1564 *
1565 * 1. packets in flight
1566 * 2. backlog
1567 * 3. prequeue
1568 * 4. receive_queue
1569 *
1570 * Each queue can be processed only if the next ones
1571 * are empty. At this point we have empty receive_queue.
1572 * But prequeue _can_ be not empty after 2nd iteration,
1573 * when we jumped to start of loop because backlog
1574 * processing added something to receive_queue.
1575 * We cannot release_sock(), because backlog contains
1576 * packets arrived _after_ prequeued ones.
1577 *
1578 * Shortly, algorithm is clear --- to process all
1579 * the queues in order. We could make it more directly,
1580 * requeueing packets from backlog to prequeue, if
1581 * is not empty. It is more elegant, but eats cycles,
1582 * unfortunately.
1583 */
1584 if (!skb_queue_empty(&tp->ucopy.prequeue))
1585 goto do_prequeue;
1586
1587 /* __ Set realtime policy in scheduler __ */
1588 }
1589
1590 #ifdef CONFIG_NET_DMA
1591 if (tp->ucopy.dma_chan)
1592 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1593 #endif
1594 if (copied >= target) {
1595 /* Do not sleep, just process backlog. */
1596 release_sock(sk);
1597 lock_sock(sk);
1598 } else
1599 sk_wait_data(sk, &timeo);
1600
1601 #ifdef CONFIG_NET_DMA
1602 tcp_service_net_dma(sk, false); /* Don't block */
1603 tp->ucopy.wakeup = 0;
1604 #endif
1605
1606 if (user_recv) {
1607 int chunk;
1608
1609 /* __ Restore normal policy in scheduler __ */
1610
1611 if ((chunk = len - tp->ucopy.len) != 0) {
1612 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1613 len -= chunk;
1614 copied += chunk;
1615 }
1616
1617 if (tp->rcv_nxt == tp->copied_seq &&
1618 !skb_queue_empty(&tp->ucopy.prequeue)) {
1619 do_prequeue:
1620 tcp_prequeue_process(sk);
1621
1622 if ((chunk = len - tp->ucopy.len) != 0) {
1623 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1624 len -= chunk;
1625 copied += chunk;
1626 }
1627 }
1628 }
1629 if ((flags & MSG_PEEK) &&
1630 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1631 if (net_ratelimit())
1632 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1633 current->comm, task_pid_nr(current));
1634 peek_seq = tp->copied_seq;
1635 }
1636 continue;
1637
1638 found_ok_skb:
1639 /* Ok so how much can we use? */
1640 used = skb->len - offset;
1641 if (len < used)
1642 used = len;
1643
1644 /* Do we have urgent data here? */
1645 if (tp->urg_data) {
1646 u32 urg_offset = tp->urg_seq - *seq;
1647 if (urg_offset < used) {
1648 if (!urg_offset) {
1649 if (!sock_flag(sk, SOCK_URGINLINE)) {
1650 ++*seq;
1651 urg_hole++;
1652 offset++;
1653 used--;
1654 if (!used)
1655 goto skip_copy;
1656 }
1657 } else
1658 used = urg_offset;
1659 }
1660 }
1661
1662 if (!(flags & MSG_TRUNC)) {
1663 #ifdef CONFIG_NET_DMA
1664 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1665 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1666
1667 if (tp->ucopy.dma_chan) {
1668 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1669 tp->ucopy.dma_chan, skb, offset,
1670 msg->msg_iov, used,
1671 tp->ucopy.pinned_list);
1672
1673 if (tp->ucopy.dma_cookie < 0) {
1674
1675 printk(KERN_ALERT "dma_cookie < 0\n");
1676
1677 /* Exception. Bailout! */
1678 if (!copied)
1679 copied = -EFAULT;
1680 break;
1681 }
1682
1683 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1684
1685 if ((offset + used) == skb->len)
1686 copied_early = 1;
1687
1688 } else
1689 #endif
1690 {
1691 err = skb_copy_datagram_iovec(skb, offset,
1692 msg->msg_iov, used);
1693 if (err) {
1694 /* Exception. Bailout! */
1695 if (!copied)
1696 copied = -EFAULT;
1697 break;
1698 }
1699 }
1700 }
1701
1702 *seq += used;
1703 copied += used;
1704 len -= used;
1705
1706 tcp_rcv_space_adjust(sk);
1707
1708 skip_copy:
1709 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1710 tp->urg_data = 0;
1711 tcp_fast_path_check(sk);
1712 }
1713 if (used + offset < skb->len)
1714 continue;
1715
1716 if (tcp_hdr(skb)->fin)
1717 goto found_fin_ok;
1718 if (!(flags & MSG_PEEK)) {
1719 sk_eat_skb(sk, skb, copied_early);
1720 copied_early = 0;
1721 }
1722 continue;
1723
1724 found_fin_ok:
1725 /* Process the FIN. */
1726 ++*seq;
1727 if (!(flags & MSG_PEEK)) {
1728 sk_eat_skb(sk, skb, copied_early);
1729 copied_early = 0;
1730 }
1731 break;
1732 } while (len > 0);
1733
1734 if (user_recv) {
1735 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1736 int chunk;
1737
1738 tp->ucopy.len = copied > 0 ? len : 0;
1739
1740 tcp_prequeue_process(sk);
1741
1742 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1743 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1744 len -= chunk;
1745 copied += chunk;
1746 }
1747 }
1748
1749 tp->ucopy.task = NULL;
1750 tp->ucopy.len = 0;
1751 }
1752
1753 #ifdef CONFIG_NET_DMA
1754 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1755 tp->ucopy.dma_chan = NULL;
1756
1757 if (tp->ucopy.pinned_list) {
1758 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1759 tp->ucopy.pinned_list = NULL;
1760 }
1761 #endif
1762
1763 /* According to UNIX98, msg_name/msg_namelen are ignored
1764 * on connected socket. I was just happy when found this 8) --ANK
1765 */
1766
1767 /* Clean up data we have read: This will do ACK frames. */
1768 tcp_cleanup_rbuf(sk, copied);
1769
1770 TCP_CHECK_TIMER(sk);
1771 release_sock(sk);
1772 return copied;
1773
1774 out:
1775 TCP_CHECK_TIMER(sk);
1776 release_sock(sk);
1777 return err;
1778
1779 recv_urg:
1780 err = tcp_recv_urg(sk, msg, len, flags);
1781 goto out;
1782 }
1783 EXPORT_SYMBOL(tcp_recvmsg);
1784
1785 void tcp_set_state(struct sock *sk, int state)
1786 {
1787 int oldstate = sk->sk_state;
1788
1789 switch (state) {
1790 case TCP_ESTABLISHED:
1791 if (oldstate != TCP_ESTABLISHED)
1792 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1793 break;
1794
1795 case TCP_CLOSE:
1796 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1797 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1798
1799 sk->sk_prot->unhash(sk);
1800 if (inet_csk(sk)->icsk_bind_hash &&
1801 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1802 inet_put_port(sk);
1803 /* fall through */
1804 default:
1805 if (oldstate == TCP_ESTABLISHED)
1806 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1807 }
1808
1809 /* Change state AFTER socket is unhashed to avoid closed
1810 * socket sitting in hash tables.
1811 */
1812 sk->sk_state = state;
1813
1814 #ifdef STATE_TRACE
1815 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1816 #endif
1817 }
1818 EXPORT_SYMBOL_GPL(tcp_set_state);
1819
1820 /*
1821 * State processing on a close. This implements the state shift for
1822 * sending our FIN frame. Note that we only send a FIN for some
1823 * states. A shutdown() may have already sent the FIN, or we may be
1824 * closed.
1825 */
1826
1827 static const unsigned char new_state[16] = {
1828 /* current state: new state: action: */
1829 /* (Invalid) */ TCP_CLOSE,
1830 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1831 /* TCP_SYN_SENT */ TCP_CLOSE,
1832 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1833 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1834 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1835 /* TCP_TIME_WAIT */ TCP_CLOSE,
1836 /* TCP_CLOSE */ TCP_CLOSE,
1837 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1838 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1839 /* TCP_LISTEN */ TCP_CLOSE,
1840 /* TCP_CLOSING */ TCP_CLOSING,
1841 };
1842
1843 static int tcp_close_state(struct sock *sk)
1844 {
1845 int next = (int)new_state[sk->sk_state];
1846 int ns = next & TCP_STATE_MASK;
1847
1848 tcp_set_state(sk, ns);
1849
1850 return next & TCP_ACTION_FIN;
1851 }
1852
1853 /*
1854 * Shutdown the sending side of a connection. Much like close except
1855 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1856 */
1857
1858 void tcp_shutdown(struct sock *sk, int how)
1859 {
1860 /* We need to grab some memory, and put together a FIN,
1861 * and then put it into the queue to be sent.
1862 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1863 */
1864 if (!(how & SEND_SHUTDOWN))
1865 return;
1866
1867 /* If we've already sent a FIN, or it's a closed state, skip this. */
1868 if ((1 << sk->sk_state) &
1869 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1870 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1871 /* Clear out any half completed packets. FIN if needed. */
1872 if (tcp_close_state(sk))
1873 tcp_send_fin(sk);
1874 }
1875 }
1876 EXPORT_SYMBOL(tcp_shutdown);
1877
1878 void tcp_close(struct sock *sk, long timeout)
1879 {
1880 struct sk_buff *skb;
1881 int data_was_unread = 0;
1882 int state;
1883
1884 lock_sock(sk);
1885 sk->sk_shutdown = SHUTDOWN_MASK;
1886
1887 if (sk->sk_state == TCP_LISTEN) {
1888 tcp_set_state(sk, TCP_CLOSE);
1889
1890 /* Special case. */
1891 inet_csk_listen_stop(sk);
1892
1893 goto adjudge_to_death;
1894 }
1895
1896 /* We need to flush the recv. buffs. We do this only on the
1897 * descriptor close, not protocol-sourced closes, because the
1898 * reader process may not have drained the data yet!
1899 */
1900 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1901 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1902 tcp_hdr(skb)->fin;
1903 data_was_unread += len;
1904 __kfree_skb(skb);
1905 }
1906
1907 sk_mem_reclaim(sk);
1908
1909 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1910 if (sk->sk_state == TCP_CLOSE)
1911 goto adjudge_to_death;
1912
1913 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1914 * data was lost. To witness the awful effects of the old behavior of
1915 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1916 * GET in an FTP client, suspend the process, wait for the client to
1917 * advertise a zero window, then kill -9 the FTP client, wheee...
1918 * Note: timeout is always zero in such a case.
1919 */
1920 if (data_was_unread) {
1921 /* Unread data was tossed, zap the connection. */
1922 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1923 tcp_set_state(sk, TCP_CLOSE);
1924 tcp_send_active_reset(sk, sk->sk_allocation);
1925 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1926 /* Check zero linger _after_ checking for unread data. */
1927 sk->sk_prot->disconnect(sk, 0);
1928 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1929 } else if (tcp_close_state(sk)) {
1930 /* We FIN if the application ate all the data before
1931 * zapping the connection.
1932 */
1933
1934 /* RED-PEN. Formally speaking, we have broken TCP state
1935 * machine. State transitions:
1936 *
1937 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1938 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1939 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1940 *
1941 * are legal only when FIN has been sent (i.e. in window),
1942 * rather than queued out of window. Purists blame.
1943 *
1944 * F.e. "RFC state" is ESTABLISHED,
1945 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1946 *
1947 * The visible declinations are that sometimes
1948 * we enter time-wait state, when it is not required really
1949 * (harmless), do not send active resets, when they are
1950 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1951 * they look as CLOSING or LAST_ACK for Linux)
1952 * Probably, I missed some more holelets.
1953 * --ANK
1954 */
1955 tcp_send_fin(sk);
1956 }
1957
1958 sk_stream_wait_close(sk, timeout);
1959
1960 adjudge_to_death:
1961 state = sk->sk_state;
1962 sock_hold(sk);
1963 sock_orphan(sk);
1964
1965 /* It is the last release_sock in its life. It will remove backlog. */
1966 release_sock(sk);
1967
1968
1969 /* Now socket is owned by kernel and we acquire BH lock
1970 to finish close. No need to check for user refs.
1971 */
1972 local_bh_disable();
1973 bh_lock_sock(sk);
1974 WARN_ON(sock_owned_by_user(sk));
1975
1976 percpu_counter_inc(sk->sk_prot->orphan_count);
1977
1978 /* Have we already been destroyed by a softirq or backlog? */
1979 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1980 goto out;
1981
1982 /* This is a (useful) BSD violating of the RFC. There is a
1983 * problem with TCP as specified in that the other end could
1984 * keep a socket open forever with no application left this end.
1985 * We use a 3 minute timeout (about the same as BSD) then kill
1986 * our end. If they send after that then tough - BUT: long enough
1987 * that we won't make the old 4*rto = almost no time - whoops
1988 * reset mistake.
1989 *
1990 * Nope, it was not mistake. It is really desired behaviour
1991 * f.e. on http servers, when such sockets are useless, but
1992 * consume significant resources. Let's do it with special
1993 * linger2 option. --ANK
1994 */
1995
1996 if (sk->sk_state == TCP_FIN_WAIT2) {
1997 struct tcp_sock *tp = tcp_sk(sk);
1998 if (tp->linger2 < 0) {
1999 tcp_set_state(sk, TCP_CLOSE);
2000 tcp_send_active_reset(sk, GFP_ATOMIC);
2001 NET_INC_STATS_BH(sock_net(sk),
2002 LINUX_MIB_TCPABORTONLINGER);
2003 } else {
2004 const int tmo = tcp_fin_time(sk);
2005
2006 if (tmo > TCP_TIMEWAIT_LEN) {
2007 inet_csk_reset_keepalive_timer(sk,
2008 tmo - TCP_TIMEWAIT_LEN);
2009 } else {
2010 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2011 goto out;
2012 }
2013 }
2014 }
2015 if (sk->sk_state != TCP_CLOSE) {
2016 sk_mem_reclaim(sk);
2017 if (tcp_too_many_orphans(sk, 0)) {
2018 if (net_ratelimit())
2019 printk(KERN_INFO "TCP: too many of orphaned "
2020 "sockets\n");
2021 tcp_set_state(sk, TCP_CLOSE);
2022 tcp_send_active_reset(sk, GFP_ATOMIC);
2023 NET_INC_STATS_BH(sock_net(sk),
2024 LINUX_MIB_TCPABORTONMEMORY);
2025 }
2026 }
2027
2028 if (sk->sk_state == TCP_CLOSE)
2029 inet_csk_destroy_sock(sk);
2030 /* Otherwise, socket is reprieved until protocol close. */
2031
2032 out:
2033 bh_unlock_sock(sk);
2034 local_bh_enable();
2035 sock_put(sk);
2036 }
2037 EXPORT_SYMBOL(tcp_close);
2038
2039 /* These states need RST on ABORT according to RFC793 */
2040
2041 static inline int tcp_need_reset(int state)
2042 {
2043 return (1 << state) &
2044 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2045 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2046 }
2047
2048 int tcp_disconnect(struct sock *sk, int flags)
2049 {
2050 struct inet_sock *inet = inet_sk(sk);
2051 struct inet_connection_sock *icsk = inet_csk(sk);
2052 struct tcp_sock *tp = tcp_sk(sk);
2053 int err = 0;
2054 int old_state = sk->sk_state;
2055
2056 if (old_state != TCP_CLOSE)
2057 tcp_set_state(sk, TCP_CLOSE);
2058
2059 /* ABORT function of RFC793 */
2060 if (old_state == TCP_LISTEN) {
2061 inet_csk_listen_stop(sk);
2062 } else if (tcp_need_reset(old_state) ||
2063 (tp->snd_nxt != tp->write_seq &&
2064 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2065 /* The last check adjusts for discrepancy of Linux wrt. RFC
2066 * states
2067 */
2068 tcp_send_active_reset(sk, gfp_any());
2069 sk->sk_err = ECONNRESET;
2070 } else if (old_state == TCP_SYN_SENT)
2071 sk->sk_err = ECONNRESET;
2072
2073 tcp_clear_xmit_timers(sk);
2074 __skb_queue_purge(&sk->sk_receive_queue);
2075 tcp_write_queue_purge(sk);
2076 __skb_queue_purge(&tp->out_of_order_queue);
2077 #ifdef CONFIG_NET_DMA
2078 __skb_queue_purge(&sk->sk_async_wait_queue);
2079 #endif
2080
2081 inet->inet_dport = 0;
2082
2083 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2084 inet_reset_saddr(sk);
2085
2086 sk->sk_shutdown = 0;
2087 sock_reset_flag(sk, SOCK_DONE);
2088 tp->srtt = 0;
2089 if ((tp->write_seq += tp->max_window + 2) == 0)
2090 tp->write_seq = 1;
2091 icsk->icsk_backoff = 0;
2092 tp->snd_cwnd = 2;
2093 icsk->icsk_probes_out = 0;
2094 tp->packets_out = 0;
2095 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2096 tp->snd_cwnd_cnt = 0;
2097 tp->bytes_acked = 0;
2098 tp->window_clamp = 0;
2099 tcp_set_ca_state(sk, TCP_CA_Open);
2100 tcp_clear_retrans(tp);
2101 inet_csk_delack_init(sk);
2102 tcp_init_send_head(sk);
2103 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2104 __sk_dst_reset(sk);
2105
2106 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2107
2108 sk->sk_error_report(sk);
2109 return err;
2110 }
2111 EXPORT_SYMBOL(tcp_disconnect);
2112
2113 /*
2114 * Socket option code for TCP.
2115 */
2116 static int do_tcp_setsockopt(struct sock *sk, int level,
2117 int optname, char __user *optval, unsigned int optlen)
2118 {
2119 struct tcp_sock *tp = tcp_sk(sk);
2120 struct inet_connection_sock *icsk = inet_csk(sk);
2121 int val;
2122 int err = 0;
2123
2124 /* These are data/string values, all the others are ints */
2125 switch (optname) {
2126 case TCP_CONGESTION: {
2127 char name[TCP_CA_NAME_MAX];
2128
2129 if (optlen < 1)
2130 return -EINVAL;
2131
2132 val = strncpy_from_user(name, optval,
2133 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2134 if (val < 0)
2135 return -EFAULT;
2136 name[val] = 0;
2137
2138 lock_sock(sk);
2139 err = tcp_set_congestion_control(sk, name);
2140 release_sock(sk);
2141 return err;
2142 }
2143 case TCP_COOKIE_TRANSACTIONS: {
2144 struct tcp_cookie_transactions ctd;
2145 struct tcp_cookie_values *cvp = NULL;
2146
2147 if (sizeof(ctd) > optlen)
2148 return -EINVAL;
2149 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2150 return -EFAULT;
2151
2152 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2153 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2154 return -EINVAL;
2155
2156 if (ctd.tcpct_cookie_desired == 0) {
2157 /* default to global value */
2158 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2159 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2160 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2161 return -EINVAL;
2162 }
2163
2164 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2165 /* Supercedes all other values */
2166 lock_sock(sk);
2167 if (tp->cookie_values != NULL) {
2168 kref_put(&tp->cookie_values->kref,
2169 tcp_cookie_values_release);
2170 tp->cookie_values = NULL;
2171 }
2172 tp->rx_opt.cookie_in_always = 0; /* false */
2173 tp->rx_opt.cookie_out_never = 1; /* true */
2174 release_sock(sk);
2175 return err;
2176 }
2177
2178 /* Allocate ancillary memory before locking.
2179 */
2180 if (ctd.tcpct_used > 0 ||
2181 (tp->cookie_values == NULL &&
2182 (sysctl_tcp_cookie_size > 0 ||
2183 ctd.tcpct_cookie_desired > 0 ||
2184 ctd.tcpct_s_data_desired > 0))) {
2185 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2186 GFP_KERNEL);
2187 if (cvp == NULL)
2188 return -ENOMEM;
2189
2190 kref_init(&cvp->kref);
2191 }
2192 lock_sock(sk);
2193 tp->rx_opt.cookie_in_always =
2194 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2195 tp->rx_opt.cookie_out_never = 0; /* false */
2196
2197 if (tp->cookie_values != NULL) {
2198 if (cvp != NULL) {
2199 /* Changed values are recorded by a changed
2200 * pointer, ensuring the cookie will differ,
2201 * without separately hashing each value later.
2202 */
2203 kref_put(&tp->cookie_values->kref,
2204 tcp_cookie_values_release);
2205 } else {
2206 cvp = tp->cookie_values;
2207 }
2208 }
2209
2210 if (cvp != NULL) {
2211 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2212
2213 if (ctd.tcpct_used > 0) {
2214 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2215 ctd.tcpct_used);
2216 cvp->s_data_desired = ctd.tcpct_used;
2217 cvp->s_data_constant = 1; /* true */
2218 } else {
2219 /* No constant payload data. */
2220 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2221 cvp->s_data_constant = 0; /* false */
2222 }
2223
2224 tp->cookie_values = cvp;
2225 }
2226 release_sock(sk);
2227 return err;
2228 }
2229 default:
2230 /* fallthru */
2231 break;
2232 }
2233
2234 if (optlen < sizeof(int))
2235 return -EINVAL;
2236
2237 if (get_user(val, (int __user *)optval))
2238 return -EFAULT;
2239
2240 lock_sock(sk);
2241
2242 switch (optname) {
2243 case TCP_MAXSEG:
2244 /* Values greater than interface MTU won't take effect. However
2245 * at the point when this call is done we typically don't yet
2246 * know which interface is going to be used */
2247 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2248 err = -EINVAL;
2249 break;
2250 }
2251 tp->rx_opt.user_mss = val;
2252 break;
2253
2254 case TCP_NODELAY:
2255 if (val) {
2256 /* TCP_NODELAY is weaker than TCP_CORK, so that
2257 * this option on corked socket is remembered, but
2258 * it is not activated until cork is cleared.
2259 *
2260 * However, when TCP_NODELAY is set we make
2261 * an explicit push, which overrides even TCP_CORK
2262 * for currently queued segments.
2263 */
2264 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2265 tcp_push_pending_frames(sk);
2266 } else {
2267 tp->nonagle &= ~TCP_NAGLE_OFF;
2268 }
2269 break;
2270
2271 case TCP_THIN_LINEAR_TIMEOUTS:
2272 if (val < 0 || val > 1)
2273 err = -EINVAL;
2274 else
2275 tp->thin_lto = val;
2276 break;
2277
2278 case TCP_THIN_DUPACK:
2279 if (val < 0 || val > 1)
2280 err = -EINVAL;
2281 else
2282 tp->thin_dupack = val;
2283 break;
2284
2285 case TCP_CORK:
2286 /* When set indicates to always queue non-full frames.
2287 * Later the user clears this option and we transmit
2288 * any pending partial frames in the queue. This is
2289 * meant to be used alongside sendfile() to get properly
2290 * filled frames when the user (for example) must write
2291 * out headers with a write() call first and then use
2292 * sendfile to send out the data parts.
2293 *
2294 * TCP_CORK can be set together with TCP_NODELAY and it is
2295 * stronger than TCP_NODELAY.
2296 */
2297 if (val) {
2298 tp->nonagle |= TCP_NAGLE_CORK;
2299 } else {
2300 tp->nonagle &= ~TCP_NAGLE_CORK;
2301 if (tp->nonagle&TCP_NAGLE_OFF)
2302 tp->nonagle |= TCP_NAGLE_PUSH;
2303 tcp_push_pending_frames(sk);
2304 }
2305 break;
2306
2307 case TCP_KEEPIDLE:
2308 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2309 err = -EINVAL;
2310 else {
2311 tp->keepalive_time = val * HZ;
2312 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2313 !((1 << sk->sk_state) &
2314 (TCPF_CLOSE | TCPF_LISTEN))) {
2315 u32 elapsed = keepalive_time_elapsed(tp);
2316 if (tp->keepalive_time > elapsed)
2317 elapsed = tp->keepalive_time - elapsed;
2318 else
2319 elapsed = 0;
2320 inet_csk_reset_keepalive_timer(sk, elapsed);
2321 }
2322 }
2323 break;
2324 case TCP_KEEPINTVL:
2325 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2326 err = -EINVAL;
2327 else
2328 tp->keepalive_intvl = val * HZ;
2329 break;
2330 case TCP_KEEPCNT:
2331 if (val < 1 || val > MAX_TCP_KEEPCNT)
2332 err = -EINVAL;
2333 else
2334 tp->keepalive_probes = val;
2335 break;
2336 case TCP_SYNCNT:
2337 if (val < 1 || val > MAX_TCP_SYNCNT)
2338 err = -EINVAL;
2339 else
2340 icsk->icsk_syn_retries = val;
2341 break;
2342
2343 case TCP_LINGER2:
2344 if (val < 0)
2345 tp->linger2 = -1;
2346 else if (val > sysctl_tcp_fin_timeout / HZ)
2347 tp->linger2 = 0;
2348 else
2349 tp->linger2 = val * HZ;
2350 break;
2351
2352 case TCP_DEFER_ACCEPT:
2353 /* Translate value in seconds to number of retransmits */
2354 icsk->icsk_accept_queue.rskq_defer_accept =
2355 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2356 TCP_RTO_MAX / HZ);
2357 break;
2358
2359 case TCP_WINDOW_CLAMP:
2360 if (!val) {
2361 if (sk->sk_state != TCP_CLOSE) {
2362 err = -EINVAL;
2363 break;
2364 }
2365 tp->window_clamp = 0;
2366 } else
2367 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2368 SOCK_MIN_RCVBUF / 2 : val;
2369 break;
2370
2371 case TCP_QUICKACK:
2372 if (!val) {
2373 icsk->icsk_ack.pingpong = 1;
2374 } else {
2375 icsk->icsk_ack.pingpong = 0;
2376 if ((1 << sk->sk_state) &
2377 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2378 inet_csk_ack_scheduled(sk)) {
2379 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2380 tcp_cleanup_rbuf(sk, 1);
2381 if (!(val & 1))
2382 icsk->icsk_ack.pingpong = 1;
2383 }
2384 }
2385 break;
2386
2387 #ifdef CONFIG_TCP_MD5SIG
2388 case TCP_MD5SIG:
2389 /* Read the IP->Key mappings from userspace */
2390 err = tp->af_specific->md5_parse(sk, optval, optlen);
2391 break;
2392 #endif
2393 case TCP_USER_TIMEOUT:
2394 /* Cap the max timeout in ms TCP will retry/retrans
2395 * before giving up and aborting (ETIMEDOUT) a connection.
2396 */
2397 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2398 break;
2399 default:
2400 err = -ENOPROTOOPT;
2401 break;
2402 }
2403
2404 release_sock(sk);
2405 return err;
2406 }
2407
2408 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2409 unsigned int optlen)
2410 {
2411 struct inet_connection_sock *icsk = inet_csk(sk);
2412
2413 if (level != SOL_TCP)
2414 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2415 optval, optlen);
2416 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2417 }
2418 EXPORT_SYMBOL(tcp_setsockopt);
2419
2420 #ifdef CONFIG_COMPAT
2421 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2422 char __user *optval, unsigned int optlen)
2423 {
2424 if (level != SOL_TCP)
2425 return inet_csk_compat_setsockopt(sk, level, optname,
2426 optval, optlen);
2427 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2428 }
2429 EXPORT_SYMBOL(compat_tcp_setsockopt);
2430 #endif
2431
2432 /* Return information about state of tcp endpoint in API format. */
2433 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2434 {
2435 struct tcp_sock *tp = tcp_sk(sk);
2436 const struct inet_connection_sock *icsk = inet_csk(sk);
2437 u32 now = tcp_time_stamp;
2438
2439 memset(info, 0, sizeof(*info));
2440
2441 info->tcpi_state = sk->sk_state;
2442 info->tcpi_ca_state = icsk->icsk_ca_state;
2443 info->tcpi_retransmits = icsk->icsk_retransmits;
2444 info->tcpi_probes = icsk->icsk_probes_out;
2445 info->tcpi_backoff = icsk->icsk_backoff;
2446
2447 if (tp->rx_opt.tstamp_ok)
2448 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2449 if (tcp_is_sack(tp))
2450 info->tcpi_options |= TCPI_OPT_SACK;
2451 if (tp->rx_opt.wscale_ok) {
2452 info->tcpi_options |= TCPI_OPT_WSCALE;
2453 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2454 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2455 }
2456
2457 if (tp->ecn_flags&TCP_ECN_OK)
2458 info->tcpi_options |= TCPI_OPT_ECN;
2459
2460 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2461 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2462 info->tcpi_snd_mss = tp->mss_cache;
2463 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2464
2465 if (sk->sk_state == TCP_LISTEN) {
2466 info->tcpi_unacked = sk->sk_ack_backlog;
2467 info->tcpi_sacked = sk->sk_max_ack_backlog;
2468 } else {
2469 info->tcpi_unacked = tp->packets_out;
2470 info->tcpi_sacked = tp->sacked_out;
2471 }
2472 info->tcpi_lost = tp->lost_out;
2473 info->tcpi_retrans = tp->retrans_out;
2474 info->tcpi_fackets = tp->fackets_out;
2475
2476 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2477 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2478 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2479
2480 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2481 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2482 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2483 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2484 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2485 info->tcpi_snd_cwnd = tp->snd_cwnd;
2486 info->tcpi_advmss = tp->advmss;
2487 info->tcpi_reordering = tp->reordering;
2488
2489 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2490 info->tcpi_rcv_space = tp->rcvq_space.space;
2491
2492 info->tcpi_total_retrans = tp->total_retrans;
2493 }
2494 EXPORT_SYMBOL_GPL(tcp_get_info);
2495
2496 static int do_tcp_getsockopt(struct sock *sk, int level,
2497 int optname, char __user *optval, int __user *optlen)
2498 {
2499 struct inet_connection_sock *icsk = inet_csk(sk);
2500 struct tcp_sock *tp = tcp_sk(sk);
2501 int val, len;
2502
2503 if (get_user(len, optlen))
2504 return -EFAULT;
2505
2506 len = min_t(unsigned int, len, sizeof(int));
2507
2508 if (len < 0)
2509 return -EINVAL;
2510
2511 switch (optname) {
2512 case TCP_MAXSEG:
2513 val = tp->mss_cache;
2514 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2515 val = tp->rx_opt.user_mss;
2516 break;
2517 case TCP_NODELAY:
2518 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2519 break;
2520 case TCP_CORK:
2521 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2522 break;
2523 case TCP_KEEPIDLE:
2524 val = keepalive_time_when(tp) / HZ;
2525 break;
2526 case TCP_KEEPINTVL:
2527 val = keepalive_intvl_when(tp) / HZ;
2528 break;
2529 case TCP_KEEPCNT:
2530 val = keepalive_probes(tp);
2531 break;
2532 case TCP_SYNCNT:
2533 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2534 break;
2535 case TCP_LINGER2:
2536 val = tp->linger2;
2537 if (val >= 0)
2538 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2539 break;
2540 case TCP_DEFER_ACCEPT:
2541 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2542 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2543 break;
2544 case TCP_WINDOW_CLAMP:
2545 val = tp->window_clamp;
2546 break;
2547 case TCP_INFO: {
2548 struct tcp_info info;
2549
2550 if (get_user(len, optlen))
2551 return -EFAULT;
2552
2553 tcp_get_info(sk, &info);
2554
2555 len = min_t(unsigned int, len, sizeof(info));
2556 if (put_user(len, optlen))
2557 return -EFAULT;
2558 if (copy_to_user(optval, &info, len))
2559 return -EFAULT;
2560 return 0;
2561 }
2562 case TCP_QUICKACK:
2563 val = !icsk->icsk_ack.pingpong;
2564 break;
2565
2566 case TCP_CONGESTION:
2567 if (get_user(len, optlen))
2568 return -EFAULT;
2569 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2570 if (put_user(len, optlen))
2571 return -EFAULT;
2572 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2573 return -EFAULT;
2574 return 0;
2575
2576 case TCP_COOKIE_TRANSACTIONS: {
2577 struct tcp_cookie_transactions ctd;
2578 struct tcp_cookie_values *cvp = tp->cookie_values;
2579
2580 if (get_user(len, optlen))
2581 return -EFAULT;
2582 if (len < sizeof(ctd))
2583 return -EINVAL;
2584
2585 memset(&ctd, 0, sizeof(ctd));
2586 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2587 TCP_COOKIE_IN_ALWAYS : 0)
2588 | (tp->rx_opt.cookie_out_never ?
2589 TCP_COOKIE_OUT_NEVER : 0);
2590
2591 if (cvp != NULL) {
2592 ctd.tcpct_flags |= (cvp->s_data_in ?
2593 TCP_S_DATA_IN : 0)
2594 | (cvp->s_data_out ?
2595 TCP_S_DATA_OUT : 0);
2596
2597 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2598 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2599
2600 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2601 cvp->cookie_pair_size);
2602 ctd.tcpct_used = cvp->cookie_pair_size;
2603 }
2604
2605 if (put_user(sizeof(ctd), optlen))
2606 return -EFAULT;
2607 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2608 return -EFAULT;
2609 return 0;
2610 }
2611 case TCP_THIN_LINEAR_TIMEOUTS:
2612 val = tp->thin_lto;
2613 break;
2614 case TCP_THIN_DUPACK:
2615 val = tp->thin_dupack;
2616 break;
2617
2618 case TCP_USER_TIMEOUT:
2619 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2620 break;
2621 default:
2622 return -ENOPROTOOPT;
2623 }
2624
2625 if (put_user(len, optlen))
2626 return -EFAULT;
2627 if (copy_to_user(optval, &val, len))
2628 return -EFAULT;
2629 return 0;
2630 }
2631
2632 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2633 int __user *optlen)
2634 {
2635 struct inet_connection_sock *icsk = inet_csk(sk);
2636
2637 if (level != SOL_TCP)
2638 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2639 optval, optlen);
2640 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2641 }
2642 EXPORT_SYMBOL(tcp_getsockopt);
2643
2644 #ifdef CONFIG_COMPAT
2645 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2646 char __user *optval, int __user *optlen)
2647 {
2648 if (level != SOL_TCP)
2649 return inet_csk_compat_getsockopt(sk, level, optname,
2650 optval, optlen);
2651 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2652 }
2653 EXPORT_SYMBOL(compat_tcp_getsockopt);
2654 #endif
2655
2656 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2657 {
2658 struct sk_buff *segs = ERR_PTR(-EINVAL);
2659 struct tcphdr *th;
2660 unsigned thlen;
2661 unsigned int seq;
2662 __be32 delta;
2663 unsigned int oldlen;
2664 unsigned int mss;
2665
2666 if (!pskb_may_pull(skb, sizeof(*th)))
2667 goto out;
2668
2669 th = tcp_hdr(skb);
2670 thlen = th->doff * 4;
2671 if (thlen < sizeof(*th))
2672 goto out;
2673
2674 if (!pskb_may_pull(skb, thlen))
2675 goto out;
2676
2677 oldlen = (u16)~skb->len;
2678 __skb_pull(skb, thlen);
2679
2680 mss = skb_shinfo(skb)->gso_size;
2681 if (unlikely(skb->len <= mss))
2682 goto out;
2683
2684 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2685 /* Packet is from an untrusted source, reset gso_segs. */
2686 int type = skb_shinfo(skb)->gso_type;
2687
2688 if (unlikely(type &
2689 ~(SKB_GSO_TCPV4 |
2690 SKB_GSO_DODGY |
2691 SKB_GSO_TCP_ECN |
2692 SKB_GSO_TCPV6 |
2693 0) ||
2694 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2695 goto out;
2696
2697 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2698
2699 segs = NULL;
2700 goto out;
2701 }
2702
2703 segs = skb_segment(skb, features);
2704 if (IS_ERR(segs))
2705 goto out;
2706
2707 delta = htonl(oldlen + (thlen + mss));
2708
2709 skb = segs;
2710 th = tcp_hdr(skb);
2711 seq = ntohl(th->seq);
2712
2713 do {
2714 th->fin = th->psh = 0;
2715
2716 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2717 (__force u32)delta));
2718 if (skb->ip_summed != CHECKSUM_PARTIAL)
2719 th->check =
2720 csum_fold(csum_partial(skb_transport_header(skb),
2721 thlen, skb->csum));
2722
2723 seq += mss;
2724 skb = skb->next;
2725 th = tcp_hdr(skb);
2726
2727 th->seq = htonl(seq);
2728 th->cwr = 0;
2729 } while (skb->next);
2730
2731 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2732 skb->data_len);
2733 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2734 (__force u32)delta));
2735 if (skb->ip_summed != CHECKSUM_PARTIAL)
2736 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2737 thlen, skb->csum));
2738
2739 out:
2740 return segs;
2741 }
2742 EXPORT_SYMBOL(tcp_tso_segment);
2743
2744 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2745 {
2746 struct sk_buff **pp = NULL;
2747 struct sk_buff *p;
2748 struct tcphdr *th;
2749 struct tcphdr *th2;
2750 unsigned int len;
2751 unsigned int thlen;
2752 __be32 flags;
2753 unsigned int mss = 1;
2754 unsigned int hlen;
2755 unsigned int off;
2756 int flush = 1;
2757 int i;
2758
2759 off = skb_gro_offset(skb);
2760 hlen = off + sizeof(*th);
2761 th = skb_gro_header_fast(skb, off);
2762 if (skb_gro_header_hard(skb, hlen)) {
2763 th = skb_gro_header_slow(skb, hlen, off);
2764 if (unlikely(!th))
2765 goto out;
2766 }
2767
2768 thlen = th->doff * 4;
2769 if (thlen < sizeof(*th))
2770 goto out;
2771
2772 hlen = off + thlen;
2773 if (skb_gro_header_hard(skb, hlen)) {
2774 th = skb_gro_header_slow(skb, hlen, off);
2775 if (unlikely(!th))
2776 goto out;
2777 }
2778
2779 skb_gro_pull(skb, thlen);
2780
2781 len = skb_gro_len(skb);
2782 flags = tcp_flag_word(th);
2783
2784 for (; (p = *head); head = &p->next) {
2785 if (!NAPI_GRO_CB(p)->same_flow)
2786 continue;
2787
2788 th2 = tcp_hdr(p);
2789
2790 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2791 NAPI_GRO_CB(p)->same_flow = 0;
2792 continue;
2793 }
2794
2795 goto found;
2796 }
2797
2798 goto out_check_final;
2799
2800 found:
2801 flush = NAPI_GRO_CB(p)->flush;
2802 flush |= (__force int)(flags & TCP_FLAG_CWR);
2803 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2804 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2805 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2806 for (i = sizeof(*th); i < thlen; i += 4)
2807 flush |= *(u32 *)((u8 *)th + i) ^
2808 *(u32 *)((u8 *)th2 + i);
2809
2810 mss = skb_shinfo(p)->gso_size;
2811
2812 flush |= (len - 1) >= mss;
2813 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2814
2815 if (flush || skb_gro_receive(head, skb)) {
2816 mss = 1;
2817 goto out_check_final;
2818 }
2819
2820 p = *head;
2821 th2 = tcp_hdr(p);
2822 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2823
2824 out_check_final:
2825 flush = len < mss;
2826 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2827 TCP_FLAG_RST | TCP_FLAG_SYN |
2828 TCP_FLAG_FIN));
2829
2830 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2831 pp = head;
2832
2833 out:
2834 NAPI_GRO_CB(skb)->flush |= flush;
2835
2836 return pp;
2837 }
2838 EXPORT_SYMBOL(tcp_gro_receive);
2839
2840 int tcp_gro_complete(struct sk_buff *skb)
2841 {
2842 struct tcphdr *th = tcp_hdr(skb);
2843
2844 skb->csum_start = skb_transport_header(skb) - skb->head;
2845 skb->csum_offset = offsetof(struct tcphdr, check);
2846 skb->ip_summed = CHECKSUM_PARTIAL;
2847
2848 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2849
2850 if (th->cwr)
2851 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2852
2853 return 0;
2854 }
2855 EXPORT_SYMBOL(tcp_gro_complete);
2856
2857 #ifdef CONFIG_TCP_MD5SIG
2858 static unsigned long tcp_md5sig_users;
2859 static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool;
2860 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2861
2862 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool)
2863 {
2864 int cpu;
2865 for_each_possible_cpu(cpu) {
2866 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2867 if (p) {
2868 if (p->md5_desc.tfm)
2869 crypto_free_hash(p->md5_desc.tfm);
2870 kfree(p);
2871 }
2872 }
2873 free_percpu(pool);
2874 }
2875
2876 void tcp_free_md5sig_pool(void)
2877 {
2878 struct tcp_md5sig_pool * __percpu *pool = NULL;
2879
2880 spin_lock_bh(&tcp_md5sig_pool_lock);
2881 if (--tcp_md5sig_users == 0) {
2882 pool = tcp_md5sig_pool;
2883 tcp_md5sig_pool = NULL;
2884 }
2885 spin_unlock_bh(&tcp_md5sig_pool_lock);
2886 if (pool)
2887 __tcp_free_md5sig_pool(pool);
2888 }
2889 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2890
2891 static struct tcp_md5sig_pool * __percpu *
2892 __tcp_alloc_md5sig_pool(struct sock *sk)
2893 {
2894 int cpu;
2895 struct tcp_md5sig_pool * __percpu *pool;
2896
2897 pool = alloc_percpu(struct tcp_md5sig_pool *);
2898 if (!pool)
2899 return NULL;
2900
2901 for_each_possible_cpu(cpu) {
2902 struct tcp_md5sig_pool *p;
2903 struct crypto_hash *hash;
2904
2905 p = kzalloc(sizeof(*p), sk->sk_allocation);
2906 if (!p)
2907 goto out_free;
2908 *per_cpu_ptr(pool, cpu) = p;
2909
2910 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2911 if (!hash || IS_ERR(hash))
2912 goto out_free;
2913
2914 p->md5_desc.tfm = hash;
2915 }
2916 return pool;
2917 out_free:
2918 __tcp_free_md5sig_pool(pool);
2919 return NULL;
2920 }
2921
2922 struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2923 {
2924 struct tcp_md5sig_pool * __percpu *pool;
2925 int alloc = 0;
2926
2927 retry:
2928 spin_lock_bh(&tcp_md5sig_pool_lock);
2929 pool = tcp_md5sig_pool;
2930 if (tcp_md5sig_users++ == 0) {
2931 alloc = 1;
2932 spin_unlock_bh(&tcp_md5sig_pool_lock);
2933 } else if (!pool) {
2934 tcp_md5sig_users--;
2935 spin_unlock_bh(&tcp_md5sig_pool_lock);
2936 cpu_relax();
2937 goto retry;
2938 } else
2939 spin_unlock_bh(&tcp_md5sig_pool_lock);
2940
2941 if (alloc) {
2942 /* we cannot hold spinlock here because this may sleep. */
2943 struct tcp_md5sig_pool * __percpu *p;
2944
2945 p = __tcp_alloc_md5sig_pool(sk);
2946 spin_lock_bh(&tcp_md5sig_pool_lock);
2947 if (!p) {
2948 tcp_md5sig_users--;
2949 spin_unlock_bh(&tcp_md5sig_pool_lock);
2950 return NULL;
2951 }
2952 pool = tcp_md5sig_pool;
2953 if (pool) {
2954 /* oops, it has already been assigned. */
2955 spin_unlock_bh(&tcp_md5sig_pool_lock);
2956 __tcp_free_md5sig_pool(p);
2957 } else {
2958 tcp_md5sig_pool = pool = p;
2959 spin_unlock_bh(&tcp_md5sig_pool_lock);
2960 }
2961 }
2962 return pool;
2963 }
2964 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2965
2966
2967 /**
2968 * tcp_get_md5sig_pool - get md5sig_pool for this user
2969 *
2970 * We use percpu structure, so if we succeed, we exit with preemption
2971 * and BH disabled, to make sure another thread or softirq handling
2972 * wont try to get same context.
2973 */
2974 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2975 {
2976 struct tcp_md5sig_pool * __percpu *p;
2977
2978 local_bh_disable();
2979
2980 spin_lock(&tcp_md5sig_pool_lock);
2981 p = tcp_md5sig_pool;
2982 if (p)
2983 tcp_md5sig_users++;
2984 spin_unlock(&tcp_md5sig_pool_lock);
2985
2986 if (p)
2987 return *this_cpu_ptr(p);
2988
2989 local_bh_enable();
2990 return NULL;
2991 }
2992 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2993
2994 void tcp_put_md5sig_pool(void)
2995 {
2996 local_bh_enable();
2997 tcp_free_md5sig_pool();
2998 }
2999 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3000
3001 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3002 struct tcphdr *th)
3003 {
3004 struct scatterlist sg;
3005 int err;
3006
3007 __sum16 old_checksum = th->check;
3008 th->check = 0;
3009 /* options aren't included in the hash */
3010 sg_init_one(&sg, th, sizeof(struct tcphdr));
3011 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
3012 th->check = old_checksum;
3013 return err;
3014 }
3015 EXPORT_SYMBOL(tcp_md5_hash_header);
3016
3017 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3018 struct sk_buff *skb, unsigned header_len)
3019 {
3020 struct scatterlist sg;
3021 const struct tcphdr *tp = tcp_hdr(skb);
3022 struct hash_desc *desc = &hp->md5_desc;
3023 unsigned i;
3024 const unsigned head_data_len = skb_headlen(skb) > header_len ?
3025 skb_headlen(skb) - header_len : 0;
3026 const struct skb_shared_info *shi = skb_shinfo(skb);
3027 struct sk_buff *frag_iter;
3028
3029 sg_init_table(&sg, 1);
3030
3031 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3032 if (crypto_hash_update(desc, &sg, head_data_len))
3033 return 1;
3034
3035 for (i = 0; i < shi->nr_frags; ++i) {
3036 const struct skb_frag_struct *f = &shi->frags[i];
3037 sg_set_page(&sg, f->page, f->size, f->page_offset);
3038 if (crypto_hash_update(desc, &sg, f->size))
3039 return 1;
3040 }
3041
3042 skb_walk_frags(skb, frag_iter)
3043 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3044 return 1;
3045
3046 return 0;
3047 }
3048 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3049
3050 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
3051 {
3052 struct scatterlist sg;
3053
3054 sg_init_one(&sg, key->key, key->keylen);
3055 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3056 }
3057 EXPORT_SYMBOL(tcp_md5_hash_key);
3058
3059 #endif
3060
3061 /**
3062 * Each Responder maintains up to two secret values concurrently for
3063 * efficient secret rollover. Each secret value has 4 states:
3064 *
3065 * Generating. (tcp_secret_generating != tcp_secret_primary)
3066 * Generates new Responder-Cookies, but not yet used for primary
3067 * verification. This is a short-term state, typically lasting only
3068 * one round trip time (RTT).
3069 *
3070 * Primary. (tcp_secret_generating == tcp_secret_primary)
3071 * Used both for generation and primary verification.
3072 *
3073 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3074 * Used for verification, until the first failure that can be
3075 * verified by the newer Generating secret. At that time, this
3076 * cookie's state is changed to Secondary, and the Generating
3077 * cookie's state is changed to Primary. This is a short-term state,
3078 * typically lasting only one round trip time (RTT).
3079 *
3080 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3081 * Used for secondary verification, after primary verification
3082 * failures. This state lasts no more than twice the Maximum Segment
3083 * Lifetime (2MSL). Then, the secret is discarded.
3084 */
3085 struct tcp_cookie_secret {
3086 /* The secret is divided into two parts. The digest part is the
3087 * equivalent of previously hashing a secret and saving the state,
3088 * and serves as an initialization vector (IV). The message part
3089 * serves as the trailing secret.
3090 */
3091 u32 secrets[COOKIE_WORKSPACE_WORDS];
3092 unsigned long expires;
3093 };
3094
3095 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3096 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3097 #define TCP_SECRET_LIFE (HZ * 600)
3098
3099 static struct tcp_cookie_secret tcp_secret_one;
3100 static struct tcp_cookie_secret tcp_secret_two;
3101
3102 /* Essentially a circular list, without dynamic allocation. */
3103 static struct tcp_cookie_secret *tcp_secret_generating;
3104 static struct tcp_cookie_secret *tcp_secret_primary;
3105 static struct tcp_cookie_secret *tcp_secret_retiring;
3106 static struct tcp_cookie_secret *tcp_secret_secondary;
3107
3108 static DEFINE_SPINLOCK(tcp_secret_locker);
3109
3110 /* Select a pseudo-random word in the cookie workspace.
3111 */
3112 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3113 {
3114 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3115 }
3116
3117 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3118 * Called in softirq context.
3119 * Returns: 0 for success.
3120 */
3121 int tcp_cookie_generator(u32 *bakery)
3122 {
3123 unsigned long jiffy = jiffies;
3124
3125 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3126 spin_lock_bh(&tcp_secret_locker);
3127 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3128 /* refreshed by another */
3129 memcpy(bakery,
3130 &tcp_secret_generating->secrets[0],
3131 COOKIE_WORKSPACE_WORDS);
3132 } else {
3133 /* still needs refreshing */
3134 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3135
3136 /* The first time, paranoia assumes that the
3137 * randomization function isn't as strong. But,
3138 * this secret initialization is delayed until
3139 * the last possible moment (packet arrival).
3140 * Although that time is observable, it is
3141 * unpredictably variable. Mash in the most
3142 * volatile clock bits available, and expire the
3143 * secret extra quickly.
3144 */
3145 if (unlikely(tcp_secret_primary->expires ==
3146 tcp_secret_secondary->expires)) {
3147 struct timespec tv;
3148
3149 getnstimeofday(&tv);
3150 bakery[COOKIE_DIGEST_WORDS+0] ^=
3151 (u32)tv.tv_nsec;
3152
3153 tcp_secret_secondary->expires = jiffy
3154 + TCP_SECRET_1MSL
3155 + (0x0f & tcp_cookie_work(bakery, 0));
3156 } else {
3157 tcp_secret_secondary->expires = jiffy
3158 + TCP_SECRET_LIFE
3159 + (0xff & tcp_cookie_work(bakery, 1));
3160 tcp_secret_primary->expires = jiffy
3161 + TCP_SECRET_2MSL
3162 + (0x1f & tcp_cookie_work(bakery, 2));
3163 }
3164 memcpy(&tcp_secret_secondary->secrets[0],
3165 bakery, COOKIE_WORKSPACE_WORDS);
3166
3167 rcu_assign_pointer(tcp_secret_generating,
3168 tcp_secret_secondary);
3169 rcu_assign_pointer(tcp_secret_retiring,
3170 tcp_secret_primary);
3171 /*
3172 * Neither call_rcu() nor synchronize_rcu() needed.
3173 * Retiring data is not freed. It is replaced after
3174 * further (locked) pointer updates, and a quiet time
3175 * (minimum 1MSL, maximum LIFE - 2MSL).
3176 */
3177 }
3178 spin_unlock_bh(&tcp_secret_locker);
3179 } else {
3180 rcu_read_lock_bh();
3181 memcpy(bakery,
3182 &rcu_dereference(tcp_secret_generating)->secrets[0],
3183 COOKIE_WORKSPACE_WORDS);
3184 rcu_read_unlock_bh();
3185 }
3186 return 0;
3187 }
3188 EXPORT_SYMBOL(tcp_cookie_generator);
3189
3190 void tcp_done(struct sock *sk)
3191 {
3192 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3193 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3194
3195 tcp_set_state(sk, TCP_CLOSE);
3196 tcp_clear_xmit_timers(sk);
3197
3198 sk->sk_shutdown = SHUTDOWN_MASK;
3199
3200 if (!sock_flag(sk, SOCK_DEAD))
3201 sk->sk_state_change(sk);
3202 else
3203 inet_csk_destroy_sock(sk);
3204 }
3205 EXPORT_SYMBOL_GPL(tcp_done);
3206
3207 extern struct tcp_congestion_ops tcp_reno;
3208
3209 static __initdata unsigned long thash_entries;
3210 static int __init set_thash_entries(char *str)
3211 {
3212 if (!str)
3213 return 0;
3214 thash_entries = simple_strtoul(str, &str, 0);
3215 return 1;
3216 }
3217 __setup("thash_entries=", set_thash_entries);
3218
3219 void __init tcp_init(void)
3220 {
3221 struct sk_buff *skb = NULL;
3222 unsigned long nr_pages, limit;
3223 int i, max_share, cnt;
3224 unsigned long jiffy = jiffies;
3225
3226 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3227
3228 percpu_counter_init(&tcp_sockets_allocated, 0);
3229 percpu_counter_init(&tcp_orphan_count, 0);
3230 tcp_hashinfo.bind_bucket_cachep =
3231 kmem_cache_create("tcp_bind_bucket",
3232 sizeof(struct inet_bind_bucket), 0,
3233 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3234
3235 /* Size and allocate the main established and bind bucket
3236 * hash tables.
3237 *
3238 * The methodology is similar to that of the buffer cache.
3239 */
3240 tcp_hashinfo.ehash =
3241 alloc_large_system_hash("TCP established",
3242 sizeof(struct inet_ehash_bucket),
3243 thash_entries,
3244 (totalram_pages >= 128 * 1024) ?
3245 13 : 15,
3246 0,
3247 NULL,
3248 &tcp_hashinfo.ehash_mask,
3249 thash_entries ? 0 : 512 * 1024);
3250 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3251 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3252 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3253 }
3254 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3255 panic("TCP: failed to alloc ehash_locks");
3256 tcp_hashinfo.bhash =
3257 alloc_large_system_hash("TCP bind",
3258 sizeof(struct inet_bind_hashbucket),
3259 tcp_hashinfo.ehash_mask + 1,
3260 (totalram_pages >= 128 * 1024) ?
3261 13 : 15,
3262 0,
3263 &tcp_hashinfo.bhash_size,
3264 NULL,
3265 64 * 1024);
3266 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3267 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3268 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3269 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3270 }
3271
3272
3273 cnt = tcp_hashinfo.ehash_mask + 1;
3274
3275 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3276 sysctl_tcp_max_orphans = cnt / 2;
3277 sysctl_max_syn_backlog = max(128, cnt / 256);
3278
3279 /* Set the pressure threshold to be a fraction of global memory that
3280 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
3281 * memory, with a floor of 128 pages.
3282 */
3283 nr_pages = totalram_pages - totalhigh_pages;
3284 limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
3285 limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
3286 limit = max(limit, 128UL);
3287 sysctl_tcp_mem[0] = limit / 4 * 3;
3288 sysctl_tcp_mem[1] = limit;
3289 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3290
3291 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3292 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3293 max_share = min(4UL*1024*1024, limit);
3294
3295 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3296 sysctl_tcp_wmem[1] = 16*1024;
3297 sysctl_tcp_wmem[2] = max(64*1024, max_share);
3298
3299 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3300 sysctl_tcp_rmem[1] = 87380;
3301 sysctl_tcp_rmem[2] = max(87380, max_share);
3302
3303 printk(KERN_INFO "TCP: Hash tables configured "
3304 "(established %u bind %u)\n",
3305 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3306
3307 tcp_register_congestion_control(&tcp_reno);
3308
3309 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3310 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3311 tcp_secret_one.expires = jiffy; /* past due */
3312 tcp_secret_two.expires = jiffy; /* past due */
3313 tcp_secret_generating = &tcp_secret_one;
3314 tcp_secret_primary = &tcp_secret_one;
3315 tcp_secret_retiring = &tcp_secret_two;
3316 tcp_secret_secondary = &tcp_secret_two;
3317 }
linux-next